Eric reported that native_sched_clock() gets miscompiled; we loose a seqcount
loop. These patches cure that.
After these patches:
native_sched_clock:
pushq %rbp #
movq %rsp, %rbp #,
... jump label ...
rdtsc
salq $32, %rdx #, tmp110
orq %rax, %rdx # low, tmp110
movq %rdx, %r10 # tmp110, _23
movq $cyc2ns, %r9 #, tmp136
.L235:
movl %gs:cyc2ns+32(%rip),%eax # cyc2ns.seq.sequence, pfo_ret__
movl %eax, %ecx # pfo_ret__, idx
andl $1, %ecx #, idx
salq $4, %rcx #, tmp116
addq %r9, %rcx # tmp136, tmp117
movq %gs:8(%rcx),%rdi # cyc2ns.data[idx_14].cyc2ns_offset, pfo_ret__
movl %gs:(%rcx),%esi # cyc2ns.data[idx_14].cyc2ns_mul, pfo_ret__
movl %gs:4(%rcx),%ecx # cyc2ns.data[idx_14].cyc2ns_shift, pfo_ret__
movl %gs:cyc2ns+32(%rip),%r8d # cyc2ns.seq.sequence, pfo_ret__
cmpl %r8d, %eax # pfo_ret__, pfo_ret__
jne .L235 #,
movl %esi, %esi # pfo_ret__, pfo_ret__
movq %rsi, %rax # pfo_ret__, tmp133
mulq %r10 # _23
shrdq %rdx, %rax # pfo_ret__,, tmp134
shrq %cl, %rdx # pfo_ret__,
testb $64, %cl #, pfo_ret__
cmovne %rdx, %rax #,, tmp134
addq %rdi, %rax # pfo_ret__, <retval>
popq %rbp #
ret
